Biology inspired growth in meta-learning
Cullen A LaKemper, Cehong Wang, Jason A. Yoder · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2022
Unlike biological systems most current neural networks models do not change their structure during learning. In most machine learning problems, it is important to choose an appropriate network structure because simple networks are likely to under-fit while complex networks are less plastic and more computationally expensive to train. To address this challenge, we introduce a dynamically growing network that starts from a small network and adds layers while training. We demonstrate an acceleration of learning using a simple dynamically growing network compared to other baseline models with fixed network structures on a meta-learning task.